Fanar-1-9B-Instruct-gguf/README.md

language, base_model, pipeline_tag

language	base_model	pipeline_tag
ar en	QCRI/Fanar-1-9B-Instruct	text-generation

Fanar-1-9B-Instruct — GGUF quantized

This repo contains multiple GGUF builds of the Arabic-English LLM Fanar-1-9B-Instruct, the instruction-tuned variant of Fanar-1-9B created by QCRI / HBKU. The base model is a 9 B-parameter continuation of gemma-2-9b trained on ≈1 T Arabic + English tokens and aligned through SFT → DPO (4.5 M / 250 K pairs). License remains Apache-2.0 and the context window is 4 096 tokens. :contentReference[oaicite:0]{index=0}

---

Available files

Bits	Format	Size (≈)
Q2_K	2-bit	3.4 GB
Q3_K_M	3-bit	4.4 GB
Q4_0 / Q4_K_M	4-bit	5.1 GB / 5.4 GB
Q5_0 / Q5_K_M	5-bit	6.1 GB / 6.3 GB
Q6_K	6-bit	8 GB*
Q8_0	8-bit	9.3 GB
F16 / F32	16 / 32-bit	17.6 GB

*value shown on the HF page is a placeholder.

---

Quick start (llama.cpp ≥ 0.2)

git clone https://github.com/ggerganov/llama.cpp && cd llama.cpp && make -j
./main -m Fanar-1-9B-Instruct.Q4_K_M.gguf -p "ما هي عاصمة قطر؟"

Python (llama-cpp-python)

from llama_cpp import Llama
llm = Llama(
    model_path="Fanar-1-9B-Instruct.Q4_K_M.gguf",
    n_ctx=4096,
    chat_format="gemma"   # Fanar follows Gemma chat template
)
print(llm.create_chat_completion(
    messages=[{"role":"user","content":"Translate 'peace' to Arabic"}]
).choices[0].message.content)

---

Credits & notes

• Original model: QCRI/Fanar-1-9B-Instruct (please consult its model card for training data, evaluation results and limitations). (Hugging Face)
• This repository only supplies GGUF conversions for efficient local inference on CPU/GPU; no weights were changed.
• Use responsibly—outputs may be inaccurate, biased, or culturally sensitive.